基于自动编码器的本色布疵点检测算法  

Grey fabric defect detection algorithm based on auto-encoder

作　　者：刘海军[1] 张莉丽 耿贵珍 朱世谊 LIU Haijun;ZHANG Lili;GENG Guizhen;ZHU Shiyi(Institute of Intelligent Information Processing,Institute of Disaster Prevention, Langfang, Hebei 065201 ,China;Institute of Ecological Environmental, Institute of Disaster Prevention, Langfang, Hebei 065201 ,China;Institute of Economics and Management, Institute of Disaster Prevention, Langfang, Hebei 065201 ,China;General Education Department, Hainan College of Vocation and Technique, Haikou, Hainan 570216,China)

机构地区：[1]防灾科技学院智能信息处理研究所,河北廊坊065201 [2]防灾科技学院生态环境学院,河北廊坊065201 [3]防灾科技学院经济管理学院,河北廊坊065201 [4]海南职业技术学院通识教育学院,海南海口570216

出　　处：《毛纺科技》2019年第9期79-83,共5页Wool Textile Journal

基　　金：河北省高等学校科学技术研究项目(ZD2019313);中央高校基本科研业务费青年教师资助计划项目(ZD2019313)

摘　　要：为解决用于本色布疵点检测的浅层机器学习方法中人工特征提取主观性强、同一种特征提取方法无法适用于不同组织结构织物的问题,采用具有特征学习功能的自动编码器神经网络对原始图像进行特征自动提取。设计了含有一个隐藏层的全连接恒等神经网络,原始数据输入该神经网络后,被隐藏层压缩,并在输出层重构,训练过程中通过优化重构层与输入层之间的误差来求解神经网络最佳系数。将训练好的自动编码器神经网络用于对原始图像进行编码压缩,经过压缩后的数据通常维数远远低于输入数据,将压缩结果作为输入图像所对应的特征向量,采用支持向量机进行分类。通过将应用自动编码器自动提取的特征与传统的PCA、HOG特征进行对比实验,结果表明,采用自动编码器自动提取的特征性能明显优于传统手工提取的特征。Nowadays the main methods in grey fabric defects detecting always come from traditional machine learning. The first step when using traditional machine learning is to extract features by hand. However, it depends on researcher′s experience to find proper features. Also, one feature may not work well to different texture structures. To solve this problem, this paper adopted auto-encoder network, which had the ability to learn features from data, to automatically extract features from original input. This paper constructed a 3-layer auto-encoder network, the input layer, the hidden layer and the output layer. Original image was straightened and put into the input layer and compressed at the hidden layer and reconstructed at the output layer. The network was trained by minimizing the error between the input image and the reconstructed image. When the training process was finished, the auto-encoder network can be used to code and compress the input image. The vector which generated during the compressed process was always short than input, that was taken as the feature. Then SVM classifier was used to classify the samples. Finally, the auto-encoder method was compared with PCA and HOG. The experimental results showed that auto-encoder precedes obviously than PCA and HOG.

关 键 词：深度学习 自动编码器 疵点检测 支持向量机 特征学习 

分 类 号：TS941.79[轻工技术与工程—服装设计与工程]